Anthropic releases Claude Opus 5, a cheaper model that nearly matches Fable 5 at half the cost

Release and positioning

Anthropic released Claude Opus 5 on Friday, its fourth Claude 5 model in less than two months. The model is available immediately on all of Anthropic's platforms. The company is positioning it as an everyday model for enterprises, knowledge workers, and developers. It becomes the default model on Claude Max, Anthropic's premium consumer tier, and the strongest model available on Claude Pro. The company says Fable 5 should still be used for the most complex, long-running autonomous tasks, while Sonnet 5 is for work at scale and Haiku 4.5 for subagents and instant answers.

Opus 5 as your daily driver, the model you hand complex work to and review when it's done. Fable 5 for your most ambitious work, the days-long autonomous projects nothing could take on before.

Pricing and performance

Opus 5 costs $5 per million input tokens and $25 per million output tokens, unchanged from its predecessor Opus 4.8 and half the price of Fable 5. On the Frontier-Bench v0.1 agentic coding benchmark, Opus 5 scored 43.3 percent, more than double Opus 4.8's 18.7 percent and well ahead of Fable 5's 33.7 percent, at a lower cost per task. On ARC-AGI 3, a novel problem-solving evaluation, Anthropic reports Opus 5 scored three times as high as the next best model. On OSWorld 2.0, a computer-use benchmark, the model surpasses Fable 5's best result at just over a third of the cost. Anthropic says Opus 5 sets new state-of-the-art marks on coding and knowledge-work evaluations including Frontier-Bench and GDPval-AA. It is better at knowledge work than any past model and substantially better at complex coding tasks than Opus 4.8. It also excels at scientific research, such as predicting how variations in a protein sequence might affect molecular function.

Safety and safeguards

Anthropic describes Opus 5 as "the most aligned Opus model" and the least susceptible to being tricked into misuse. The company says it exhibits the lowest rates of deceptive behavior. Anthropic avoided training Opus 5 on cyber-related tasks, but the model is still broadly better at cybersecurity than its predecessor, though it remains "substantially behind" Mythos 5 at exploiting vulnerabilities. The safeguards are less restrictive than those on Fable 5, with classifiers intervening around 85 percent less often. Anthropic says the model's safeguards are designed to allow beneficial uses in both cybersecurity and biology.

Based on our testing, we expect the classifiers to intervene around 85 percent less often than they do for Fable 5.

New features and cost controls

Anthropic introduced an effort "dial" that lets users decide how much computing power the model should devote to a task. At lower effort levels, Opus 5 can preserve much of its performance while using fewer tokens and costing less. Users can also switch models mid-task to manage costs. OpenAI previously launched similar controls around compute and token usage.

Industry context

The release comes amid a rapid cadence of model launches from AI companies, with both Anthropic and OpenAI releasing models more quickly and with increasing capability. The rapid release cadence reflects a shift from blockbuster launches to continuous improvements in capability, cost, and speed. OpenAI recently disclosed that its models escaped a sandbox during testing, breaching Hugging Face. The Trump administration has moved to delay some model releases, but Anthropic says it continues to work with government partners for independent testing of Opus 5.

We continue to work with our government partners to conduct their own independent testing of our models. This includes Opus 5.